An Intelligent Decision Support System for the Treatment of Patients Receiving Ventricular Assist Device Support

Transcription

1 Original Artiles 1 An Intelligent Deision Support System for the Treatment of Patients Reeiving Ventriular Assist Devie Support E. C. Karvounis 1,2 ; M. G. Tsipouras 1,2 ; A. T. Tzallas 1,2 ; N. S. Katertsidis 1,2 ; K. Stefanou 1,2 ; Y. Goletsis 1,3 ; M. Frigerio 4 ; A. Verde 4 ; R. Caruso 5 ; B. Meyns 6 ; J. Terrovitis 7 ; M. G. Trivella 5 ; D. I. Fotiadis 1,2 1 Biomedial Researh Institute-FORTH, Ioannina, Greee; 2 Unit of Medial Tehnology and Intelligent Information Systems, Dept. of Material Siene and Engineering, University of Ioannina, Ioannina, Greee; 3 Department of Eonomis, University of Ioannina, Ioannina, Greee; 4 Cardiology 2 Heart Failure and Transplantation Division, A. De Gasperis Cardiothorai and Vasular Department, Niguarda Ca Granda Hospital, Milan, Italy; 5 CNR Institute of Clinial Physiology, Pisa, Italy; 6 Department of Cardia Surgery, University Hospital Leuven, Leuven, Belgium; 7 3rd Cardiology Department, University of Athens, Shool of Mediine, Athens, Greee Keywords Heart failure, ventriular assist devies, deision support system Summary Bakground: Heart failure (HF) is affeting millions of people every year and it is haraterized by impaired ventriular performane, exerise intolerane and shortened life expetany. Despite signifiant advanements in drug therapy, mortality of the disease remains exessively high, as heart transplant Corespondene to: Dimitrios I. Fotiadis Unit of Medial Tehnology and Intelligent Information Systems Dept. of Material Siene and Engineering University of Ioannina PO Box Ioannina Greee fotiadis@.uoi.gr Methods Inf Med 2014; 53: doi: /ME reeived: April 29, 2013 aepted: January 1, 2014 prepublished: remains the gold standard treatment for endstage HF when no ontraindiations subsist. Traditionally, implanted Ventriular Assist Devies (VADs) have been employed in order to provide irulatory support to patients who annot survive the waiting time to transplantation, reduing the workload imposed on the heart. In many ases that proess ould reover its ontratility performane. Objetives: The SensorART platform fouses on the management and remote treatment of patients suffering from HF. It provides an interoperable, extendable and VAD-independent solution, whih inorporates various hardware and software omponents in a holisti approah, in order to improve the quality of the patients treatment and the workflow of the speialists. This paper fouses on the desription and analysis of Speialist s Deision Support System (SDSS), an innovative omponent of the SensorART platform. Methods: The SDSS is a Web-based tool that assists speialists on designing the therapy plan for their patients before and after VAD implantation, analyzing patients data, extrating new knowledge, and making informative deisions. Results: SDSS offers support to medial and VAD experts through the different phases of VAD therapy, inorporating several tools overing all related fields; Statistis, Assoiation Rules, Monitoring, Treatment, Weaning, Speed and Sution Detetion. Conlusions: SDSS and its modules have been tested in a number of patients and the results are enouraging. 1. Introdution Heart failure (HF) an be defined as an abnormality of ardia struture or funtion leading to failure of the heart to deliver oxygen at a rate ommensurate with the requirements of the metabolizing tissues, despite normal filling pressures (or only at the expense of inreased filling pressures) [1]. Common auses of heart failure inlude myoardial infartion and other forms of ishemi heart disease, hypertension, valvular heart disease and ardio - myopathy. It is a ommon, ostly, disabling and potentially deadly ondition. In developed ountries, around 2% of adults suffer from heart failure, but in those over the age of 65, this rate reahes up to 6 10% [2]. For these reasons, together with the diffiulty of having a suffiient number of heart donors, it is reognized that the de- Shattauer 2014 Methods Inf Med 2/2014

2 2 vie-based therapeuti approahes assume an inreasing and important role in the treatment of HF patients, not only as bridge to transplant, but also as destination therapy. Additionally, in reent years, there is important evidene that patients assisted by Ventriular Assist Devies (VADs) an potentially reover the natural heart funtions. VAD is a mehanial pump used to support heart funtion and blood flow in people who have weakened hearts. In this framework, the SensorART platform fouses on the management and remote treatment of patients suffering from heart failure by offering an interoperable, extendable and VAD-independent (i.e. working with different VAD types/brands) solution, whih inorporates hardware and software omponents in a holisti approah, in order to improve the quality of the patients treatment and the workflow of the speialists [3]. In addition, SensorART enables speialists to better understand the patientdevie interations, and get an insight into new knowledge. The only deision support omponent of the SensorART platform is the Speialist s Deision Support System (SDSS) [4]. The SensorART platform in general and SDSS in partiular omes to fill-in the gap of data analysis and related deision support in a rather new medial field where although an abundane of data is reated, there is no systemati way to store, aess and analyse it. Even more, in suh patient data, hidden knowledge usually resides; still there are neither available tools to mine suh data, nor related databases. Even statistial analysis and hypotheses testing are not supported by speialised tools and researhers have to reate manually input files from medial reords. Also, there are no systems apable of simultaneous monitoring and analysis of patient measurements. Suh a tool ould serve as a prognosti/alerting tool in everyday patient management and ould signifiantly improve patient treatment by early identifying patient risk. Patient risk is urrently estimated by generalised soring systems whih do not onsider individual patient profiles. The VAD performane, ustomizable throughout pump speed seletion, is mainly set empirially following a trialerror proedure without the support of intelligent proess that uses the patient profile to elaborate the suggested speed. In that framework, as all medial DSSs that use eletroni linial data and provide patient-speifi information to liniians with the aim of reduing medial errors [5], the proposed SDSS offers support towards three main diretions: i) Analysis of patient data, hypothesis testing and extration of hidden assoiations among patient variables. ii) Support on treatment deisions by providing estimation of adverse events risk and identifiation of patients that ould be onsidered as andidates for weaning from VAD, and iii) Support on optimal setting of VAD speed based on patient profiles and estimated ardia output obtained through a speially built VADirulatory system simulator [3]. Sution detetion is also employed towards this diretion. In order to reate and integrate our SDSS into the linial pratise, the proposed SDSS should at pro-ative and produe reminders, warnings, ask questions, analyse trends based on the data of the eletroni reord and on up-to-date medial knowledge while trained by the outome of their deisions [6]. The proposed SDSS provides a holisti approah by supporting medial deisions and ations in the whole patient life-yle. Before VAD implantation, using the SDSS the speialist is able to design the therapy plan by assessing patient status and the risk of adverse events assoiated to the implantation. This plan ould even inlude the deision whether to implant a VAD or not. After the implantation, the speialist an use the SDSS to get a suggestion on the optimal VAD speed, an monitor patient and VAD parameters and assess the probability for an adverse event the following day. Eventually, the speialist an assess improvements in patient status that ould denote adequate ardia reovery, whih in turn ould lead to weaning from the VAD. Speialist s researh is also supported though hypotheses testing on a speially designed repository built with hetero - geneous data oming multiple soures (laboratories, implantable and wearable sensors). In this paper we present the funtionalities of the SDSS and its sub-omponents and we demonstrate their appliation on real patient data. The authors have pre - sented an initial SDSS design in [3]; the urrent paper desribes in detail the final implementation, inluding evaluation results on real patient datasets. The paper is organized as follows: In Setion 2 an extended state of the art of related DSSs and methodologies in the field of VADs is presented. In Setion 3 tools and omponents of the SDSS are presented in detail, while, in Setion 4, the obtained results on a real patient dataset are provided. Finally, extensions and onluding remarks on the ap - pliation of SDSS tools are disussed in Setion Related Work Current linial pratie related to treatment support of left VAD (LVAD) patients is haraterized by none to limited appliation of information tehnology and intelligent deision support systems. In the literature only researh works that support individual aspets of the problem have been referred. Analysis of linial studies often neessitates omplex proesses and multiple graphial representation of the results. An exploratory data analysis inludes a variety of tehniques, in order to disover hidden relationships, suh as patterns or lusters in a dataset, identifiation of most important parameters, et. Most professional software pakages (IBM SPSS, Mathematia, SAS, et) are either only ommerially available or hard to use, espeially if one aims to generate or ustomize a huge number of similar graphial outputs. In addition, most statistis programs fore users to import data from a database before they an use them, while in many ases there is no diret onnetion to the database. Assoiation rule mining has been applied in order to extrat interesting relations between variables in medial databases. A Priori [7] is a well-known method using support-based pruning to systematially ontrol the exponential growth of andidate itemsets. Doddi et al [8] have investigated the appliation of assoiation rules in medial data. Hu [9] utilized the assoiation rules tehnique (using an Im- Methods Inf Med 2/2014 Shattauer 2014

3 3 proved A Priori algorithm) to ondut data mining on 285 ases of breast disease patients to reate rules between tumor reurrene and other attributes suh as age and tumor size. Sriraam et al [10] used a Malaysian hospital dataset in order to deide on kidney dialysis treatment using assoiation rules. Immamura et al [11] utilized an assoiation rules tehnique on 477 patients in order to investigate the three most useful linial findings for hroni diseases. Lee et al. [12] have proposed an assoiation rule mining method on 1,247 young Korean adults to extrat patterns related to aute myoardial infartion. Still, most ommerial DSSs fous mainly on extrating statistial measurements from a given patient database and not on extrating new knowl - edge. For VAD treated patients, several risk sores and assessment tools have been presented in the literature [13 30]. Some of them are mentioned here. The Heart Failure Survival Sore (HFSS) [13] has been proposed for patient seletion for LVAD support based on the estimation for expeted survival during the next 1 3 years. It has been developed using statistial likelihood analysis. In the same ontext, the Seattle Heart Failure Model (SHFM) [14] and the Randomized Evaluation of Mehanial Assistane in Treatment of Chroni Heart Failure trial (REMATCH) [15] have been also employed. The Interageny Registry for Mehanially Assisted Cirulatory Support registry (INTER- MACS) [16] has been used for patient lassifiation in risk groups, interval analysis [17], and timing of implant assessment [18]. Also, patient lassifiation related to the risk of developing other diseases when undergoing LVAD implantation has been addressed with the Model for End-Stage Liver Disease (MELD) [19]. In addition, the RVF risk sore (RVFRS) pre-operative tool [27] and the pre-operative RV stroke work index (RVSWI) [28] have also been presented in the literature. Still, most of the aforementioned studies are based on statistial analysis tehniques only, while [25] and [30] employ data mining tehniques. Current linial pratie of LVAD patients seems to be laking advaned visualization tools that an be used for monitoring day-to-day patient data. Suh tools an prove to be signifiant assistants to patient monitoring, ideally linked to prognosti systems that an predit adverse events. The value of day-to-day data reorded in a patient diary is demonstrated in several studies presented in the literature, for both heart failure patients or for patients suffering from other hroni diseases and onditions. Kirhner et al [32] proposed a heart failure patient monitoring system with implantable defibrillators. White et al. [33] performed diary data analysis for Heart Failure symptom reognition. Hayes et al [34] inluded walking program diary data to study the effets of exerise training on the quality of life in LVAD patients. As, in some ases, LVAD devies an lead to heart reovery at least this being the vision for the future- a hard deision to be taken is patient weaning. A few models have been presented in the literature. Santelies et al. [35] proposed a model derived from retrospetive medial experiene, through a series of strutured interviews and questionnaires of 11 members of the multidisiplinary Artifiial Heart Program at the University of Pittsburgh medial enter. Dandel et al. [36] analyzed data on ardia morphology and funtion olleted before VAD implantation, ehoardiographi parameters reorded during off-pump trials, duration of HF before implantation, and stability of reovery before and early after VAD removal. To assess the preditability of post-weaning outome without heart transplantation or other VAD implantation and to identify risk fators for post-weaning HF reurrene, the authors used the data olleted before and during VAD support. Birks et al. [37] proposed a set of minimum riteria with the LVAD at 6000 rpm for 15 minutes for explantation. In the ase of parameter improvement, the ombination therapy was ontinued until the maximum improvement had been ahieved in eah patient. All above weaning models propose a set of simple risp rules in order to identify weaning andidate patients and the final outome of the related models is a yes/no suggestion without any quantifiation of the strength of this suggestion. In the ase of Pump Speed Seletion, a few sensor-less methods for determining the pump speed were developed by using pump variables suh as urrent, voltage and speed [38 40]. These methods were based on the observation that pressure aross, and flow through, an LVAD an be inferred or estimated from the eletrial urrent and frequeny of the pump s motor. Several other researhers have adopted a similar approah without the use of implantable sensors [41 43]. Control shemes to keep the average pressure aross the pump (or between the aorta and the left ventrile) onstant have also been developed [44 46]. These an provide a pump speed in the safe zone or optimal zone for a ertain systemi vasular resist - ane. Other ontrol approahes, suh as using the heart rate as input have been reported [47]. Chen presented an improved method that inorporates the heart rate and the systemi vasular resistane (SVR), and responds to the physiologial hanges of the body instantaneously, based on the baroreflex and the built in ardiovasular regulation system [48]. MConahy shows the onstraints on ardia output, left atrial pressure, and arterial pressure [49]. However, the proposed methodologies for pump speed seletion present signifiant drawbaks like: they are reliable only in a relatively narrow range of pump variables, some methods do not take into aount the hange in SVR, et. Conerning detetion of sution events in the available related signals (i.e. pump flow or pump urrent signals), several approahes have been proposed to evaluate it [50 53]. These approahes are based on the empirial observation of ertain variables. Thus, some sution indies are based on time-domain features [54 57], frequeny-domain features [55 57], and time-frequeny-domain features [55 57]. Among them, there exist methods whih extrat features from the pump flow signal being one of very few signals that an be easily measured and use powerful pattern reognition algorithms to lassify the signal into different pump states [57 61]. However, in the majority of the above desribed approahes, a large number of features are employed. Also, sution detetion algorithms are used only in LVAD ontrollers [62]. Shattauer 2014 Methods Inf Med 2/2014

4 4 3. Material and Methods The SensorART platform ontinuously monitors and evaluates patient-devie interations and optimizes the heart unloading and support. In this diretion, the proposed SDSS enhanes the SensorART platform with advaned tools, thus enabling the realization of an intelligent VAD implementation. By failitating suh an approah, the SDSS aims at providing a better understanding of the patient ondition and progress in order for the speialists to personalize and optimize the orresponding treatment support, determine the apaity of the natural heart to develop major or minor delivery aording to the assisting time, identify risk fators and reovery mehanisms in order to reognize potential outomes et. Figure 1 presents the role of the SDSS within the overall SensorART platform. The SDSS resides of the server side, aesses the SensorART repository where patient data are ontinuously stored, and ommuniates with the VAD Simulation Platform (in order to aquire ustomized simulation sessions). The SensorART repository itself is EUROMACS (www. euromas.org) ompliant so as to ensure interoperability; a speial wrapper appliation ensures HL7 ompatibility, as well. Data are generated on the lient side either diretly by the VAD itself, by implantable sensors and by wearable sensors (Continua alliane ertified). An Autoregulation Unit used for VAD ontrol and data olletion, manages the data whih, in turn, are sent through a portable devie (e.g. a tablet) (running the Patients Monitoring Appliation) to the data repository. The SDSS and the Speialist s Monitoring Appliation aess this repository. They form a resoureful environment that provides valuable outomes (patient ondition /parameters, data analyses, risk fators, suggestions, estimations, et.) to the speialists (i.e. liniians, researhers, et.). The SDSS offers support to medial and VAD experts through seven tools for Assoiation Rules extration, Statistis analysis, Treatment deision Support, Monitoring, Weaning deision support, VAD Speed Seletion and Sution Detetion respetively. The SDSS tools ommuniate with the following system parts: i) the SDSS User Interfae that inludes a set of pages/forms and ontrols allowing advaned interativity with the users of the SDSS, ii) the VAD Simulation Platform providing a hybrid ventriular and irulatory model allowing speialists to simulate the behaviour of a patient s irulatory system using different VAD types and funtional parameters [63], iii) an R environment for statistial omputing, and most important iv) the Sensor- ART repository that provides storage and retrieval funtionalities for all the data used in the SensorART platform and was developed to support the management of data that are used by the orresponding tools. The SersorART repository was designed and implemented with the following data entities: users, user groups, alerts and notifiations, sensor measurements, VAD measurements, laboratory measurements, patient profiles et. Various sreenshots of the platform are presented in Figure 2 Figure 1 A general overview of the SensorART Platform Methods Inf Med 2/2014 Shattauer 2014

5 5 while all seven SDSS tools are desribed in detail in the following paragraphs. 3.1 Assoiation Rules Tool Using the Assoiation Rules Tool, the speialist an extrat assoiation rules, linking different patient variables based on patient data over a speifi observation period. In partiular, the Speialist an be assisted in analysis and researh, by using data mining tehniques in order to disover assoiations among different variables and disover new knowledge. Assoiations are in the form of Rules: IF (variable 1 > value 1) AND (variable 2 <= value 2) AND THEN (variable 3 > value 3) AND (variable n > value n). By using this tool, various hypotheses on relationships among variables an be now examined (if onfirmed) by existing data. The speialist an selet the variables that will be inluded in the IF and the THEN parts, the patients that will be inluded and the time period (start and end date) of the analysis. Furthermore, he an filter the extrated rules by setting support and onfidene thresholds. Rules together with their onfidene and support values are presented to the speialist. Moreover, in order to aquire as muh meaningful results as possible, a double filtering approah is followed by (i) ignoring features that have a high perentage of missing values in the database and (ii) ignoring reords for whih a high number of feature values is missing. The speialist an define the relevant thresholds. The A Priori algorithm [7] was hosen as the rule mining algorithm, beause of its robustness and ability to solve large problems using small omputational power and time. The A Priori algorithm attempts to find subsets whih are ommon to at least a minimum number of the itemsets, given an initial set of itemsets. The algorithm an find assoiations between different sets of data and the sets of rules (the algorithm s outome) implies how often items are ontained in sets of data. It should be mentioned that, although several different approahes have been presented in the literature, A Priori remains the benhmark algorithm of the assoiation rule mining field, while most of the other algorithms improve the omputational time for mining the assoiation rules from the dataset, still mainly produing the same set of rules [7]. The A Priori algorithm in general uses breadth-first searh and a Hash tree struture to keep trak of the andidate item sets. Then, it generates these andidate item sets of length N from the initial itemsets of length N 1. Afterwards, the andidates whih have an infrequent sub pattern are pruned. Finally, the ultimate andidate set ontains all frequent-length itemsets. The problem of assoiation rule mining is defined as: Let I = i 1, i 2,..., i d be the set of all items in a dataset, and T = t 1, t 2,..., t N be the set of all transations. Eah transation has a unique identifier and ontains a set of items, alled an itemset. An itemset with k items is alled a k-itemset. An important property of an itemset is its support ount, whih refers to the number of transations whih ontain this partiular itemset. An assoiation rule is a onditional impliation among itemsets, of the form X Þ Y where X and Y are disjoint itemsets, i.e. X Ç Y = Æ. The strength of the rule an be measured in terms of its support and on - fidene. The rule has support s, if s % of transations inlude all the items in both X and Y, and onfidene if % of transations ontaining also ontain Y. Support is an important measure, beause a rule that has very low support may our simply by hane. Confidene, on the other hand, measures the reliability of the inferene made by a rule. Hene, the data mining task for assoiation rules an be broken into two steps. The first step onsists of finding all frequent itemsets, i.e. itemsets that our in the database with a ertain user speified frequeny, alled minimum support. The seond step onsists of forming the rules among the frequent itemsets. Initially, the data and the assigned support and onfidene values are read. Then, the data are preproessed by a parsing tool, and the preproessed data are used as input to the A Priori algorithm and exports the results to a file. The parsing tool operates as follows: taking into aount the perentage of the non-missing values that the user has filled in, the algorithm ounts the number of missing values per variable and if the sum of the missing values exeeds the desired user perentage, then the whole variable is not taken into aount when the A Priori rule mining algorithm runs, produing results only for the variables that fulfill the user defined riteria. 3.2 Statistis Tool The multitude of data generated in the medial field requires the adoption of in - telligent analysis tehniques that allows speialists to summarize and present their knowledge, get insight into the data, test hypotheses, draw onlusions and diretly interat with all the available information. Thus, the Statistis Tool provides speialists with instruments for: i) Analyzing and interpreting large patient data diretly from the SensorART repository through powerful statistial tehniques and Kaplan-Meier analysis, ii) Examining the results of previous therapeuti regimens and obtaining quantitative explanations of the observations through relevant statistial tests and iii) Generating effiient reports with intelligent data visualization. Using the tool the medial expert has multiple options, inluding the seletion of population of interest, to work only with data from speifi visits, to hoose the variables of interest (e.g. heart related, sensor related, laboratory related) and the method of interest from a omplete set of basi and advaned features overing both linial and researh needs of linial partners (basi statistismean, variane, std, quantile, length, paired T-test, unpaired T-test, F-test, χ 2 tests and Kaplan-Meier analysis). The Statistis Tool is based on the R environment ( for its bak-end funtionalities exept for the Kaplan-Maier omponent. The latter allows speialists to perform survival analysis and generate a Kaplan-Meier plot. In linial studies, the researhers are usually interested in the time until the patients in a study present a speifi event or endpoint. In the ontext of SensorART, these were defined as death, erebral bleeding, gastrointestinal bleeding, ishemi stroke, TIA (transient ishemi attak) and thromboemboli events. The Kaplan-Meier tool is based on a standard algorithm for determing, diretly from the SensorART repositories, the fration of patients living Shattauer 2014 Methods Inf Med 2/2014

6 6 Figure 2 Sreenshots from (a) Assoiation Rules Tool, (b) Statistis Tool, () Treatment Support Tool and (d) Weaning Tool Methods Inf Med 2/2014 Shattauer 2014

7 7 up to the ourrene of a speifi event. The starting time is always onsidered the time of VAD implantation and thus the start of treatment. 3.3 Treatment Support Tool The Treatment Support Tool supports the speialists on the most suitable treatment plan aording to the ondition/phase of the patients (stabilized linial state, normal phase in home onditions, worsening phase and/or reautisation phase). Its funtionality inludes the alulation of several aknowledged risk sores along with the predition of alternative treatments outome with respet to adverse events appearane. More speifially, the Treatment Support Tool provides two funtionalities, alulation of known risk sores and treatment predition based on risk for adverse event appearane. In the first ase, several known sores of survival are inorporated into the tool to allow rapid deisions and to foresee possible ompliations after VAD implant. All risk sores are alulated for similar and partially overlapped but not idential objetives inluding predition of: i) Expeted survival/the risk of death on medial therapy, ii) Expeted survival/the risk of death after LVAD implantation and iii) Probability of speifi ompliations (e.g. right ventriular failure) after LVAD implantation. Unfortunately, some risk fators for death without operation are also assoiated with worse postoperative survival and/or higher probability of ompliations, making diffiult to define the risk/benefit profile for individual patients. In this ontext, four risk sores are inluded in the Treatment Support Tool: the Heart Failure Survival Sore (HFSS), whih provides an estimation of expeted 1-year survival [13], the Seattle Heart Failure Model (SHFM), whih provides an estimation of expeted 1-, 2- and 3-year patient survival [14], the Model for End-Stage Liver Disease (MELD), whih estimates risk fro multisystem malfuntions and post-operative ompliations [19], the RVF risk sore (RVFRS), whih estimates risk for right ventriular failure [27]. Conerning risk predition based on adverse events appearane, this modality assesses the risk of adverse events in the ase of LVAD implantation. It has been developed by applying mahine learning tehniques in an annotated dataset. In model building, several widely known lassifiation methodologies have been tested: Naïve Bayes lassifier (NB), k-nearest neighbor (knn), Deision trees (DT), Random forests (RF), Multilayer pereptron (MLP) neural networks and Support Vetor Mahines (SVMs), using ommonly used parameters, proposed in the literature: NB lassifier was implemented by modeling numeri values using normal distributions. knn was implemented with k = 3 and based on the Eulidean distane. DT were implemented using the C4.5 algorithm, using the pessimisti error rate based method (sub-tree replaement) for pruning, with onfidene fator 0.25 and 2 as minimum instanes in a leaf. MLP arhiteture was implemented with 1 hidden layer, 0.3 learning rate and 500 maximum number of epohs. RFs were based on 10 deision trees. SVM was implemented with polynomial kernel. Based on the obtained results (presented in Setion 4) and the fat that the medial experts preferred a mehanism that an offer interpretability of the lassifiation outome, the DT have been finally seleted. 3.4 Monitoring Tool The Monitoring Tool offers monitoring of day-to-day LVAD and patient parameters and their assoiation with the appearane of speifi adverse events. The monitoring parameters used are: Pump flow, Pump speed, Pump index, Pump power, Temperature, Systoli blood pressure, Diastoli blood pressure, Pulses, Weight, INR and Antioagulant treatment. Those parameters an ome diretly from the patient or, in the future implementation of the Sensor- ART platform, an be automatially obtained through a set of implantable and other sensors. In addition, several adverse events are also reorded inluding bleeding, arrhythmia, heart failure, thrombo - embolism (major), thromboembolism (minor) and pump thrombosis. The patterns that were used to develop the prognosti model were extrated from the dataset. Again, NB lassifier, knn, DT, RF and MLP lassifiers were tested, while DT have been finally seleted based on the obtained results (Setion 4) and the fat that the medial experts preferred the idea of a transparent deision mehanism against a blak-box approah (suh as e.g. in the ase of MLPs). 3.5 Weaning Tool The Weaning Tool an be used to identify the most appropriate andidates for weaning from the VAD support. As already mentioned in Setion 2, all proposed approahes reported in the literature for supporting optimal ventriular assist devie weaning make use only of strit rules. In our proposed tool, a Fuzzy Knowledge Subsystem is inorporated in order mainly to be more flexible on the deision boundaries and loser to the human logi ompared to lassial binary (risp) logi. More speifially, the Weaning tool enables speialists to reate and modify expert rules for weaning, in the form of omprehensive and personalized IF-THEN rules. The tool ombines expert knowledge with fuzzy analysis, in order to support the speialists on the weaning deision, i.e. the seletion of patients with adequate ardia reovery that may be removed from the VAD therapy. The tool is based on the risp engine and the fuzzy knowledge subsystem [64], in order to reate initial knowledgebased models and then transform them into fuzzy models. All patient data, suh as patient information (e.g. demographis, medial history, mediation), VAD parameters, sensor measurements, laboratory measurements and linial evaluations, are used as inputs into the fuzzy models, in order to provide the status of eah patient (andidate for weaning or not). Shattauer 2014 Methods Inf Med 2/2014

8 8 Conerning the Fuzzy Knowledge Subsystem, being the most innovative part of the Weaning tool, the main idea is to produe a fuzzy model for a speifi domain of appliation based on initial knowledge for this area provided by domain experts. This knowledge is formatted in a set of initial risp rules, i.e. a olletion of IF THEN rules. More speifially, eah risp lassifiation rule r i (x, θ i ) is expressed as: r i (x, θ i ): d i (x, θ i ) y i, where r i is the i-th risp rule, x is a feature vetor omprised from a number of features a j, θ i is a vetor of thresholds, d i is the risp rule s preondition, ontaining a onjuntion of feature tests, and y i is the predited lass. Preondition is defined as: where tion, with Figure 3 (1) is the risp membership fun- and opj = {=,, <, >,, } and defined as an be seen in Figure 3. The symbol denotes the binary AND operator. All risp rules from the medial rule set that have as onsequent the same lass are ombined in a single lass rule, defined as: (3) Symbol denotes the binary OR operator. Based on the above, a risp model an be defined as: (4) where M is the risp model, Θ is the threshold vetor ontaining all thresholds used in the model, Φ is a deision funtion and N is the number of lasses in the problem, and thus the number of lass rules R. The risp model M an be transformed into an equivalent fuzzy model M f as follows: The risp membership funtion g is replaed by a fuzzy one g f. The binary AND and OR operators are replaed by T norm and S norm funtions, respetively. The deision funtion Φ is replaed by a defuzzyfiation funtion Φ f. Based on these hanges, the fuzzy model is defined as: (5) where, θ f is the parameter vetor ontaining all parameters used in the fuzzy model and R k f is the fuzzy lass rules, defined as: Figure 4 tool [63] (a) Speed seletion flowhart, (b) Internal onnetivity of the VAD-Heart Simulation platform with the sution detetion and speed seletion Methods Inf Med 2/2014 Shattauer 2014

9 9 where r i f is the fuzzy rule defined as: (6) [7] The obtained fuzzy model M f is optimized with respet to Θ f parameters. For this purpose an objetive funtion must be defined using a training dataset (suh as the square error funtion), and then minimized with respet to a Θ f using a loal or global optimization tehnique. Loal optimization tehniques start from an initial point and result to the respetive loal minimum while global optimization tehniques attempt to find all loal minima (whih are potentially global) of the objetive funtion inside a bounded set. The desribed methodology for fuzzy expert systems reation is based on the fuzzyfiation of an initial (risp) medial rule set. Thus, the final fuzzy expert system is the initial rule set with more flexible boundaries and optimized based on a training set. The methodology an be applied to any given rule set. In the framework of the Weaning Tool, the risp and fuzzy rule engines have been developed for two relevant models in the literature i.e. the Dandel et al. [36] set of rules and Birks et al. [37] flow hart. Although, the original versions of the models inlude only risp rules, fuzzy models have also been developed sine the definition and parameter set of the fuzzy model allows it to be extremely more flexible and thus being able to ope with the omplexity of the respetive medial deision. 3.6 Speed Seletion Tool Setting the pump speed of the VAD is an important parameter in order to ahieve the optimal ardia output, and ensure the patient s quality of life. In a ritial are setting, the desired operating point of the VAD may be determined by a speialist or tehnial personnel and adjusted to provide more or less ardia output depending on the status of the patient. However, when Table 1 No Feature Age Features in the dataset for the treatment support tool INTERMACS profile Platelets Hemoglobin Hematorit White blood ells Right atrial Pressure Pulmonary Artery Pressure max Pulmonary Artery Pressure min Pulmonary Artery Pressure mean Pulmonary Capillary Wedge Cardia Index mean ± std ± (median) ± ± ± ± ± ± ± ± ± ± 0.47 No Feature Reversible PH Heart Rate PA International Normalized Ratio Bilirubine Creatinine Urea Na+ Model for End-Stage Liver Disease (MELD) MELD UNOS MELD U+Age Aspartate Anotransferase Inotropes mean ± std 1 (median) ± ± ± ± ± ± ± ± ± ± ± ± 0.87 the patient leaves the ritial are setting, a liniian is no longer readily available and the devie must provide adequate ardia output to sustain the patient s level of ativity without linial supervision. In addition, as the patient reovers and regains his strength, a hange in the VAD speed might be required. Cardia output an be inreased by inreasing the pump speed. However, two important onstraints should be taken into aount regarding the pump speed. First, if the speed is too low, blood may regurgitate from the aorta to the left ventrile through the pump (bakflow). Seond, if the pump speed is too high, the pump will attempt to draw more blood from the ventrile than is available, resulting in the sution phenomenon. A solution to maximize ardia output while operating the pump at a safe speed is to operate the pump at a speed just below sution. The drawbak to this solution is that a pump speed that does not ause sution may still have adverse effets on other physiologial parameters. In this framework, the proposed Speed Seletion Tool has been developed enabling the speialists to analyze simulation sessions from the VAD Simulation platform, investigate the potential effet on important hemodynami variables (suh as ardia output and arterial pressure) and determine a pump speed that provides adequate ardia output for the patient to maintain his urrent level of ativity. In this framework, we designed a flowhart ( Figure 4a) for the speed seletion proess urrently followed after the operation. The Left Atrial Pressure (LAP) hek is substitute for eho examination, whih is used to hek if the aorti valve is opening orretly. As shown above in the flowhart, the automated identifiation of VAD pump speed depends on the automated detetion of sution events in the pump flow signal. In this framework, a tool has been developed (Sution Detetion Tool), onneted with the Simulation Platform, in order to provide to the speialists a powerful assistant in their attempt to analyze data from simulation sessions, identify different pump states and possible issues regarding to the sution phenomenon. The proess is shown graphially in Figure 4b. The Sution Detetion Tool proesses simulation sessions from the Simulation Platform, allowing speialists to simulate the be - Shattauer 2014 Methods Inf Med 2/2014

10 10 Table 2 Treatment Support Tool Classifiation Results for Sensitivity (Sens), Positive Preditive Value (PPV) and Auray (A), for Naïve Bayes (NB), k-nearest neighbor (knn), Deision trees (DT), Random forests (RF), Multilayer pereptron (MLP) neural networks and Support Vetor Mahines (SVMs) lassifiers (a) Initial dataset (D1) (b) Initial dataset with replaed missing values (D2) Sens (%) PPV (%) A (%) Sens (%) PPV (%) A (%) NB NB knn knn DT DT RF RF MLP MLP SVM SVM () Resampled dataset (D3) (d) Resampled dataset with replaed missing values (D4) Sens (%) PPV (%) A (%) Sens (%) PPV (%) A (%) NB NB knn knn DT DT RF RF , MLP MLP SVM SVM haviour of a patient s irulatory system, using different VAD types and funtional parameters in order to be supported to the speed deision. 3.7 Sution Detetion Tool Initially, an interonneted upload mehanism allows the submission of simulation sessions from the VAD Simulation Platform. The dataset requested is omposed of ardia and irulatory variables, and other parameters important to obtain the speifi objet s (patient or animal) patho-physiologial ondition. The operator, adjusting the irulatory model parameters is able to obtain the desired hemodynami status. Some of the data transmitted from SDSS to VHSP are: patient s weight, ardia output, mean entral venous pressure, duration of simulation et, seleted aording to linial partner s suggestions. Next, the simulated pump flow signal is returned bak to SDSS, together with the following traes: Left Atrial Pressure, Left Ventriular pres - sure, Aorti Pressure, VAD Flow, heart rate, Total Cardia Output and pump speed in RPM. A new sution detetor algorithm has been developed based on the detetion of the sudden dereases in signal s baseline. The methodology is based on online estimation of a Gaussian Mixture Model (GMM) with two mixtures orresponding to non-sution & sution lasses. More speifially, the proposed methodology is onsisted of three steps: i) signal window - ing, ii) GMM based lassifiation and iii) GMM parameter adaptation. In our GMM model the only time varying parameter that was seleted is the mean of the non-sution mixture [64]. Reduing degrees of freedom in the model ensures higher reliability. 4. Datasets and Results In order to evaluate eah proposed tool, related datasets were used. In the ase of Treatment Support Tool and espeially of the risk of adverse events predition, sample data from 49 patients have been used. The variables reorded are mentioned in Table 1. The re-hospitalizations for all patients for the first year and follow-up data have been used to determine the ourrene of adverse events: among 49 patients treated with VADs; 35 had no relevant adverse events; 3 had bleeding episodes and 11 died. The dataset have been obtained from the Heart Failure and Transplantation Division, A. De Gasperis Cardiothorai & vasular Department, Niguarda Ca Granda Hospital, Milan, Italy. Although being very ommon, infetions of the entry site of the driveline were not onsidered as relevant adverse events, and thus were exluded from the study. The dataset inluded several missing values thus replaement of missing values has been applied using the 3-Nearest Neighbors tehnique. As the number of the prototypes per ategory is unbalaned, resampling from the normal distribution of the minority lass proedure has been employed, i.e. a normal distribution was alulated for eah feature, using all patterns belonging to the same lass (the minority lass that needs to be resampled). Then, N random values were generated from eah alulated distribution, with N being the number of additional samples for the lass, and thus N additional patterns belonging to this lass were generated. In addition, a feature sele- Methods Inf Med 2/2014 Shattauer 2014

11 11 Table 3 Monitoring Tool Classifiation Results for Sensitivity (Sens), Positive Preditive Value (PPV) and Auray (A), for Naïve Bayes (NB), k-nearest neighbor (knn), Deision trees (DT), Random forests (RF), Multilayer pereptron (MLP) neural networks and Support Vetor Mahines (SVMs) lassifiers (a) Initial dataset (D1) (b) Resampled dataset (D2) Sens (%) PPV (%) A (%) Sens (%) PPV (%) A (%) NB NB knn knn DT DT RF RF MLP MLP tion, using exhaustive searh and Chisquare statistis, was applied in order to selet the most informative features for this problem, resulting to the subset of features: Hemoglobin, Hematorit, Right Atrial Pressure, ardia Index, International Normalized Ratio and Aspartate Anotrans - ferase. As already mentioned in Setion 2, NB, 3NN, DT using the C4.5 algorithm, MLP neural networks and RF lassifiation methodologies have been tested. Evaluation was performed for i) the initial dataset (D1), ii) the initial dataset with replaed missing values (D2), iii) the resampled dataset (D3) and iv) resampled dataset with replaed missing values (D4). The leaveone-patient-out ross-validation tehnique was employed. The respetive onfusion matries (m ii ) were obtained, while the metris of lassifiation auray (A), sensitivity for eah lass (Sens i ) and average positive preditive value (PPV) for eah lass (PPV i ) for eah lass, are alulated: (8) Based on the above, the average sensitivity for all lasses (Sens) and average positive preditive value (PPV) for all lasses are alulated. All results are presented in Table 2. In order to develop the prognosti proedure that assesses the risk of an adverse event appearane in the next day (Monitoring Tool), based on the reorded data for the last three days, data from six patients were inluded. The dataset have been obtained from the Department of Cardia Surgery, University Hospital Leuven, Leuven, Belgium. The patterns that were used to develop the prognosti model were extrated from this dataset. In order to extrat a pattern, four onseutive days should exist in the reording and in this ase the data from the three days are used as features and the event from the fourth day as the annotation. Thus, the dataset is formulated as: D = {X, T }where X being a 33xN matrix with eah row being one data pattern of the dataset with 33 features (i.e. the 11 variables for three onseutive days), and T being the annotation of the respe - tive pattern. Using this proedure, a dataset with 1026 patterns has been reated, inluding patterns overing three lasses, being normal (1022), ventriular tahyardia (3) and heart failure (1). Initially, sine the number of the patterns per ategory is unbalaned, the same resampling proedure (resampling from the normal distribution of the minority lass) was applied to reate a balaned dataset. The lassifiation methodologies that were employed inlude NB lassifier, knn, C4.5, RF and MLP neural networks, using the parameters already mentioned previously. Evaluation was performed using: i) the 10-fold stratified ross validation method and ii) the initial dataset (before the resampling) and the respetive onfusion matries were obtained, while metris suh as lassifiation auray and sensitivity/positive preditive value per lass are alulated. The obtained results are shown in Table 3. In the framework of the Weaning Tool and the optimization of the fuzzy model, data from 11 patients are used, olleted from the 3rd Cardiology Department, Shool of Mediine of the University of Athens in Greee from 2004 to Almost 70 parameters are reorded. Four (4) of the patients were suessfully weaned from VAD. Reordings were made in three main patient onditions, i) on pump, ii) off pump for 5 mins, and iii) off pump for 15 min. The set of rules presented by Birks et al. [37] were used as input to the above desribed methodology and the results are presented using the above desribed dataset. The on-pump data and the off pump for 5 mins data were used for optimization. These inluded 9 patterns for weaning and 23 patterns for non-weaning. The training dataset is formulated as: D max = {X, T} where X being an 6x32 matrix with eah row being one data pattern of the dataset (i.e. X i = x, with i = 1,..., 32 ), and T being an 2x32 matrix with eah row being the annotation of the respetive pattern (i.e. T i = t with t = 1,..., 32). A target vetor for a speifi data pattern (t) is defined as: (9) The optimization funtion is based on the sum of square errors funtion: (10) A loal optimization tehnique has been used. As mentioned above, by setting: θ i,b f = θ i, θ i,a f = 1 it is: M f (x, Θ f ) Shattauer 2014 Methods Inf Med 2/2014

12 12 Table 4 Weaning Tool Confusion Matries for the Initial (Θ f max) and the Optimized (Θ f* ) Fuzzy Models Initial fuzzy model Θ f initial annotation Optimized fuzzy model Θ f* annotation wean no wean wean no wean fuzzy model wean 4 0 fuzzy model wean 10 0 no wean 6 22 no wean 0 22 M (x, Θ ). Thus, the initial starting point Θ f initial is defined by setting θ i,a f = 2 and θ i,b f = θ i, i.e.: Θ f initial = {2,60,2,50,2,45,2,12,2,2.8,2,16} In order to restrain the loal optimiza - tion from drifting far from the Θ f initial (and thus maintaining the properties of the initial medial rule set) upper and lower bound for eah one of the optimization parameters (Θ f ) are set so as to limit the optimization proedure inside this searh area. These bounds are set as: θ i,a f Î [0,5] and θ i,b f Î [0.9θ i, 1.1θ i ]. The optimization pro edure results to a loal minimum: Θ f * = {0.5, 59, 1, 51, 0.17, 45, 0.5, 13.5, 0.5, 2.8, 2.5, 16}. For Θ f initial and Θ f * the obtained onfusion matries are presented in Table 4 while square error, sensitivity, speifiity and lassifiation auray results in Table 5. Table 5 Weaning Tool Classifiation Results for Sensitivity (Sens), Positive Preditive Value (PPV) and Auray (A), for the Initial (Θ f max) and the Optimized (Θ f* ) Fuzzy Models Square error Sensitivity (%) Speifiity (%) Classifiation auray (%) No-Sution Sution Sensitivity Speifiity Auray DATASET I Θ f initial GMM No Sution Θ f* GMM Sution Finally, onerning the Sution Detetion Tool, two different datasets, provided by the Institute of Bioybernetis and Biomedial Engineering of Polish Aademy of Sienes (IBBE-PAS) and annotated be medial experts, are used in order to test our methodology: i. 10 pump flow signals with sution events approximately 46 minutes in total duration are olleted from the VAD Simulation Platform whih enables the speialists to simulate the behaviour of a patient s irulatory system with onneted a real assist devie (e.g. nonpulsatile blood pump) [64]. ii. 26 pump flow signals approximately 20 hours are produed from a numerial simulator, simulating different medial ases with predefined pathologies. A large number of medial realisti ases with patient pathologies were defined and determined by medial dotors [64]. The onfusion matries, sensitivities, speifiities and the total auray are given in Table 6 and Table for the two datasets, respetively. 5. Disussion In reent years, the use of VADs for the treatment of end stage heart failure has Table 6 Results from 10 signals with approximately 46 minutes (dataset I) and from 26 signals with approximately 20 hours (dataset II), where GMM online estimation was applied [63]. DATASET II GMM No Sution GMM Sution steadily inreased. This widespread appliation has inreased researh in the field, however, based on the literature review, deision support systems onerning VAD patients have not been proposed over the past years; only researh works that support individually aspets of the problem have been presented. In this ontext, the SDSS is innovative, supporting medial and VAD experts through the different phases of VAD therapy. This paper fouses on the presentation of the SDSS, the main deision support omponent of the Sensor- ART system. SDSS assists speialists on designing the best treatment plan for their patients before and after VAD implanta - tion, analysing patients data, testing hy - potheses, extrating new knowledge and making informative deisions. The proposed SDSS Tools have been integrated in a unified environment, overing all the speialist s requirements and providing a unique seletion of features. More speifially, all Tools are web-based, aessible through an intuitive graphial interfae allowing seured aess to available data. In addition, different types of data and detailed reports are provided to the medial experts from the majority of the Tools. More important, the proposed system doesn t address just a single issue but assists the speialists on remote monitoring and VAD ontrol and reeive information on patient status. It enables the design of the most proper treatment plan for every patient before and after VAD implantation and enables the analysis of patients data for the extration of new knowledge and making informative deisions. Finally, the SDSS enables the potential to translate valuable expert knowledge into standardized, personalized and optimized VAD therapy. Summarizing, the proposed system presents several advantages: Methods Inf Med 2/2014 Shattauer 2014